Arbor adds tree-search memory to autonomous optimization agents

Arbor proposes a tree-search cognition layer for autonomous agents operating in large, stateful action spaces.

Instead of treating each optimization attempt as isolated, Arbor maintains a scored tree of hypotheses that evolves as agents measure results. Failures are not discarded; they reshape the next branches of exploration.

The paper validates the approach on full-stack LLM inference optimization, a setting where performance depends on coordinated changes across models, kernels, serving, and application code.